def eval_setfit()

in scripts/setfit/run_fewshot_multilingual.py [0:0]

• 50 lines of code
• 7 McCabe index (conditional complexity)

def eval_setfit(train_data, test_data, model, loss_class, num_epochs, metric):
    x_train = train_data["text"]
    y_train = train_data["label"]

    x_test = test_data["text"]
    y_test = test_data["label"]

    if loss_class is None:
        return compute_metrics(x_train, y_train, x_test, y_test, metric)

    # sentence-transformers adaptation
    batch_size = args.batch_size
    if loss_class in [
        losses.BatchAllTripletLoss,
        losses.BatchHardTripletLoss,
        losses.BatchSemiHardTripletLoss,
        losses.BatchHardSoftMarginTripletLoss,
        SupConLoss,
    ]:
        train_examples = [InputExample(texts=[text], label=label) for text, label in zip(x_train, y_train)]
        train_data_sampler = SentenceLabelDataset(train_examples)

        batch_size = min(args.batch_size, len(train_data_sampler))
        train_dataloader = DataLoader(train_data_sampler, batch_size=batch_size, drop_last=True)

        if loss_class is losses.BatchHardSoftMarginTripletLoss:
            train_loss = loss_class(
                model=model,
                distance_metric=BatchHardTripletLossDistanceFunction.cosine_distance,
            )
        elif loss_class is SupConLoss:
            train_loss = loss_class(model=model)
        else:
            train_loss = loss_class(
                model=model,
                distance_metric=BatchHardTripletLossDistanceFunction.cosine_distance,
                margin=0.25,
            )

        train_steps = len(train_dataloader) * num_epochs
    else:
        train_examples = []
        for _ in range(num_epochs):
            train_examples = sentence_pairs_generation(np.array(x_train), np.array(y_train), train_examples)

        train_dataloader = DataLoader(train_examples, shuffle=True, batch_size=batch_size)
        train_loss = loss_class(model)
        train_steps = len(train_dataloader)

    print(f"{len(x_train)} train samples in total, {train_steps} train steps with batch size {batch_size}")

    warmup_steps = math.ceil(train_steps * 0.1)
    model.fit(
        train_objectives=[(train_dataloader, train_loss)],
        epochs=1,
        steps_per_epoch=train_steps,
        warmup_steps=warmup_steps,
        show_progress_bar=False,
    )

    return compute_metrics(x_train, y_train, x_test, y_test, metric)